synthetic-data-generate-eval

Synthetic data validation grades generated run outputs across six dimensions to ensure demo-quality fitness. It reads actual generated records to catch schema-conformant but obviously synthetic datasets that would fail domain expert review.

Out-of-chain data plausibility checking evaluates whether generated records look believable to domain experts. It catches datasets that technically conform to the form schema but contain obviously synthetic or garbage values unfit for stakeholder demos.

Grading a synthetic data run involves evaluating six weighted dimensions: record count health, form schema coverage, out-of-chain data plausibility, warning honesty, manifest provenance, and operator next steps to produce a formal quality verdict.

Schema-conformant synthetic data can still fail validation. Hard fitness gates automatically fail runs with obviously synthetic or garbage data even when all schema conformance requirements are met, ensuring only believable data passes.

Manifest provenance validation checks the metadata trail of synthetic data generation runs. It confirms the generated dataset has proper provenance records, contributing to the overall demo-quality fitness verdict.

Synthetic data evaluation is needed during ACE Phase 7 Plan B Stage 4 to validate labs-side synthetic generation for Connect opportunities. It ensures enough visits are produced and named FLWs correctly round-trip with their promised archetypes.